About the author
Lukas Alexander is ChurnZero’s vice-president of customer success. He has over a decade of experience successfully building and scaling global customer success teams, and elevating Net Revenue Retention (NRR) year over year.
**
Quick Summary: AI is only as good as what you give it. Understand the three inputs your AI is missing to make AI work in customer success.
When a new CSM joins your team, you don’t hand them a laptop and a customer list and expect results.
You onboard them.
So, now that AI is becoming more capable every day of acting like a digital teammate, why aren’t we treating it in a similar way?
The conversation about AI in customer success tends to run in one direction. What can AI do for my team? How much can I automate? How much time can we save? Until we put the same amount of thought into what AI needs from us, however, we can’t make AI work in customer success in the transformative way we hope.
Recently, I joined Carlos Quezada, VP of global services and success at Gigamon, for a ChurnZero webinar on how to scale AI-powered customer success without losing the human touch. (Watch the full conversation.) The question came up time and again.
It’s particularly urgent now because, historically, CS has had a structural data deficit compared to sales and marketing teams. Sales has had Salesforce for decades. Marketing has had HubSpot. While this is changing quickly, the data maturity gap is real, and you need to close it for your AI to deliver the best outcomes systematically.
What it takes to make AI work in customer success.
Ultimately, AI is only as good as the data, context and directions it has. Right now, too many CS teams are feeding it stale or incomplete information, then wondering why the outputs don’t land.
Until you fix it, you’re not running an AI-powered CS program; it’s more like a faster version of what you had before.
Note: At ChurnZero, this approach shaped how we built our AI agents for customer success teams. You’ll find details at the end.
1. AI needs you to tell it what to look for.
Your AI can process everything in your tech stack. However, it can’t identify which information would change the quality of a customer conversation, because it doesn’t know where CSMs are walking in blind.
Think about the last business review you sat in on. Did your CSM enter with the full picture, or were they filling gaps with instinct during the call?
The gap you observed is where AI should come in, to a) look for signals and b) act on them.
Try this exercise. Map the moments where your CSMs routinely lack what they need: pre-QBR, pre-renewal, post-onboarding handoff, 30 days after a support escalation.
For each scenario, define what intelligence is missing and what useful action that intelligence could trigger. Could it be a call with the executive sponsor, for example? A specific talk track? A change in position?
You can almost consider these moments and their corresponding actions as your AI’s job description. Your CS platform handles the structured layer: usage trends, health scores, renewal timing. The AI layer handles the relationship signals: sentiment, stakeholder engagement, and what topics have gone quiet. Most teams are strong on the first and weak on the second. That’s the gap your CSMs are filling with gut feel.
2. AI needs you to tell it when to pause.
Most AI tools keep running until you tell them not to. They might not know a stakeholder just left, or that the energy has shifted in an account, which means they don’t know when automated engagement is about to fall flat due to changed conditions.
This isn’t a technology problem but a definition problem in which, if you don’t define the conditions under which AI and automation pause, your system continues to do what it’s doing.
In our webinar, I referred to the outcome as broadcasting: one-way digital engagement that doesn’t have a human escalation layer. It isn’t how CS is supposed to work.
The solution is to build an escalation logic layer of defined “human triggers”, specific and evidence-backed, that tell your system to pause the digital track and route the account to a human.
To be confidently actionable, human triggers need to be observable, not inferred. “Customer seems disengaged” isn’t observable. “Executive sponsor hasn’t attended the last two scheduled calls, and sentiment has trended negative over 45 days” is.
Build your escalation conditions at that level of specificity, then map them to the right response, not just a flag. Does this condition trigger a CSM call? An executive reach-out? A specific talk track? The condition and the response belong together.
Now, adjust your rules to make them appropriate by customer size, product complexity, or however else you segment your service levels.
3. The data layer you need to make AI work in customer success.
“They’re three months from renewal and still logging in.” For years, CS teams treated product usage data like this as the signal that settled everything. Of course, we know a customer can be highly active in your product and still churn because the risk is in the relationship.
What do those signs look like? A champion stops being candid, or their answers get shorter. Keywords that indicate dissatisfaction show up in communications. No usage dashboard catches those signals, which means your AI can’t either, without the data layer it needs.
That data layer is what we call the relationship score.
You likely have a product usage health score. Keep it. However, you need a second layer alongside it: champion sentiment over time, meeting attendance by the right stakeholders, executive sponsor engagement, recency of real conversations on topics that matter.
These scores belong side by side in every account review. You’re looking for divergence. When usage is green and the relationship score is declining, that gap is your most important signal—because by the time churn shows up in the product, the relationship has been gone for weeks.
How ChurnZero AI is built to close the gap.
Giving your AI the information it needs doesn’t have to be a manual task. In fact, the three needs we’ve discussed above have already been addressed in the way we built ChurnZero AI.
For your intelligence backlog.
Data enrichment agents like Archetype, Pulse, Vibes, and Intel run continuously to keep pre-call intelligence current. Archetype tracks who drives decisions. Pulse monitors engagement trends. Vibes surfaces sentiment shifts, while Intel gathers strategic context. Together, they’ll build the brief your CSMs should have before every conversation.
For your human trigger library.
ChurnZero’s AI Signals layer detects engagement shifts, early risk indicators, and sentiment changes, feeding them directly into your health scores and segmentation logic. This means that your human trigger library is built into the system rather than something you have to remember. The Herald agent works the positive side, identifying advocates before your CSMs might think to look.
For relationship intelligence.
Engagement AI—the first AI-powered relationship scoring capability in any CS platform—analyzes sentiment, relationship dynamics, and conversation topics across every customer interaction. Recent updates exclude automated journey activity, so the signal reflects genuine human engagement only.
